Given the impact I think 9.2 is the earliest we could do this.

@kares @enebo @mkristian Any opinions on this?

I remember the author (A. Sund??????) of the original (and I do not think is our) JSR223 impl made some decisions about what would persist across calls. He did most of the original bindings for most of the languages (Sun employee). I was fairly sure we copied those decisions.

If I ignore that then the obvious reason this is a bad idea is it will pin old objects over time if you keep adding new lvars. If your script is ('foo = 2 unless foo') then you likely will either a) be exploiting this feature or b) going off in the weeds since you did not realize foo had been used by another call previously. With that said, there is also the added confusion that the engine keeps constants/methods/globals loaded. So that is inconsistent. If I remember @Ivars will also persist across invocations? Of all the things which could persist I think local variables are probably the most likely to be unexpected. They also have no API for de-allocating the slot in the scope. You can uninitialize if you use 'local_variables' to get a list and set to nil but the scope table will just grow forever.

With my long rambling personal opinion stated if all the other languages do this then perhaps we should too. We will be exchanging issue reports on being inconsistent with leaking memory (and changed behavior). Since we have probably had ~5 jsd223 bugs in the last 10 years we can probably go either way :)

We should also ask @yokolet what she thinks. She may remember more of the history of JSR223.

@enebo I am currently leaning toward just making this change. I have no idea how many people are using JRuby + JSR223, but I can't imagine there's many. Most folks who want to embed JRuby want the richer API, and that's what we have usually told people asking about embedding.

If you agree this would be simple to drop into 9.1.16, and see if anyone hates it.

The change is made, and only for 9.2+ because of a regression it can cause...

The old behavior was "global", which meant that local variables "put" into the engine went into $variable style names. This meant they were persistent, but this did not match the behavior of other language engines that simply use normal local variable names.

By switching the "persistent" we gain the persistent local variables, but lose the global variable. This will break code that expected the global variable.

Given the fact that we've had at least two JSR-223 engine users report that variables do not persist like they expected, we are making the change for 9.2. It can be switched back to "global" or to some other mode by setting the JVM property org.jruby.embed.localvariable.behavior=global.